INFRASTRUCTURE
them needs high throughput inference and massive training runs, all with a low enough latency for real-time decisions.”
Bringing AI into the physical world requires modeling 3D space directly rather than processing flat data.
Cormac highlights that spatial AI demands a shift in how data is structured from the ground up:“ Building deep learning infrastructure for real-world physics and 3D space starts with treating geometry as the primary data type – a first-class citizen so to speak – rather than approaching the world purely as a 2D computer vision problem. Real-world physics is unforgiving, so making the leap from 2D perception to 3D spatial intelligence requires infrastructure designed for the physical world from the outset.”
Thermal boundaries Scaling deep learning compute comes with grappling with the physical realities of both heat and power limits.
“ The density demands of AI compute have changed the game,” Colm explains.“ Air cooling can’ t keep up with what modern AI hardware requires, so at AWS we have invested heavily in liquid cooling and ground-up data centre redesign. The latest architecture delivers 12 % more compute power while cutting mechanical energy consumption by up to 46 % during peak cooling.” While the hyperscaler AWS is re-engineering megawatt facilities, field-deployed sensors on Hexagon’ s side face a very different thermal wall.
Cormac highlights the architectural trade-offs when processing physical data on handheld and autonomous hardware:“ One of the biggest challenges is that sensors are advancing faster than the edge hardware available to process everything they capture. LiDAR and imaging systems can generate increasingly rich data about the physical world, but a handheld scanner or a battery-powered robot still has finite compute, memory, energy and thermal capacity. That mismatch is the real bottleneck, not whether a model can theoretically understand a scene, but whether it can do so within those physical constraints.”
Breaking bottlenecks High-performance neural networks are only as fast as the pipelines feeding them. On the cloud side, keeping pace requires rethinking network architecture entirely.
“ The key is to avoid moving or processing more data than you actually need”
Cormac Bonner AI Program Director Hexagon
100 October 2026